Qualcomm Cloud AI 100
AI inference accelerator for data centres and edge devices
The Qualcomm Cloud AI 100 is an artificial intelligence inference card designed for servers and edge hardware. It uses a custom system on a chip with sixteen AI cores to deliver high performance per watt in data centre and low-power environments.
- Uses sixteen scalar 4-way VLIW AI cores
- Supports 25W edge and 75W data centre modes
- Built on a 7nm manufacturing process
- Connects via PCIe Gen4 x8 interface
- Includes open-source model efficiency toolkit
What it is
The Qualcomm Cloud AI 100 is a hardware accelerator built specifically for artificial intelligence inference tasks. Unlike chips designed for training neural networks, this device focuses on running existing models efficiently. Qualcomm developed the system on a chip to serve both large data centres and smaller edge deployments. The company leveraged intellectual property from its mobile division to create a solution that prioritises performance per watt. It connects to host systems via a PCIe Gen4 x8 interface, allowing it to fit into standard server slots.
How it came about
Qualcomm introduced the Cloud AI 100 at the Hot Chips 33 conference. This launch represented a continued effort by the company to enter the data centre market, following the discontinuation of its Centriq 2400 server processor. The project aimed to adapt low-power AI acceleration technology originally developed for mobile devices. By targeting higher-value segments with lower volume requirements, Qualcomm sought to establish a presence in cloud and edge infrastructure. The presentation highlighted the transition of mobile IP into higher-power inference cards.
Inside the chip
The core of the chip features sixteen Qualcomm AI cores linked by a network on chip interconnect. Each AI core uses a scalar four-way very long instruction word architecture. This design includes vector and tensor units that support lower precision calculations, which are common in inference workloads. A parallelising compiler distributes computing tasks across these cores to maximise efficiency. The system also supports peer-to-peer communication through a PCIe switch, enabling multiple cards to interact without burdening the central processor. Four LPDDR4X memory landings provide local storage for model data.
How it was made
The chip is manufactured using a seven-nanometre process node. This advanced manufacturing technique allows Qualcomm to adjust clock speeds and voltages dynamically. By lowering these parameters, the accelerator can achieve better performance per watt in power-constrained environments. The design supports two thermal design power points: seventy-five watts for data centre servers and twenty-five watts for edge devices. This flexibility enables the same silicon to serve different market segments effectively.
Where it went
The accelerator is deployed in data centre servers and edge computing devices. In one example configuration, it pairs with a Snapdragon 865 processor for edge applications. For server use, it fits into specialised chassis such as the Gigabyte G292-Z43. This server model features side pods with multiple PCIe x8 slots, allowing up to sixteen accelerator cards in a single dual-socket AMD EPYC system. The hardware targets workloads where latency and energy efficiency are critical, such as real-time image recognition.
Versions and family
The product line includes two primary form factors to suit different ecosystem needs. One variant uses the M.2 standard, while another adopts the newer EDSFF format. These physical designs allow integration into various server architectures. The company also provides software tools, including an open-source AI Model Efficiency Toolkit. This toolkit helps developers compress models and quantise data, reportedly improving performance by fifteen percent with minimal accuracy loss on benchmarks like ResNet-50.
Why it matters
This chip represents Qualcomm's strategy to leverage its mobile expertise in the data centre sector. While the company does not hold a dominant position in server hardware compared to its mobile success, the Cloud AI 100 offers an alternative for inference tasks. It competes in a space where PCIe connectivity is more prevalent than proprietary links like NVLink. The focus on open-source software tools aims to reduce the gap in developer support compared to established competitors.
Questions people ask about Qualcomm Cloud AI 100
Is the Cloud AI 100 used for training AI models?
No, the Qualcomm Cloud AI 100 is designed specifically for inference. It runs existing artificial intelligence models rather than training new ones. This focus allows the hardware to optimise for latency and power efficiency during prediction tasks.
What type of memory does the chip use?
The system on a chip includes four landings for LPDDR4X memory. This mobile-grade memory standard helps maintain low power consumption while providing sufficient bandwidth for inference workloads on the sixteen AI cores.
How does the chip handle multiple accelerators?
It supports peer-to-peer communication enabled over a PCIe switch. This allows multiple Cloud AI 100 cards to exchange data directly without routing everything through the main CPU, which is useful in servers with many installed accelerators.
Good reading and sources
There is no English Wikipedia article on Qualcomm Cloud AI 100 yet, so the facts on this page come from the sources listed above: servethehome.com/qualcomm-cloud-ai-100-ai-inference-card-at-hot-chips-. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.